Apple opens its model framework, and lets Claude in
Under the WWDC noise about Siri, Apple made a quieter move with a longer tail. It's open-sourcing the Foundation Models framework, the on-device layer behind Apple Intelligence, later this summer, and it opened that framework to third-party models. The story resurfaced to the top of Hacker News this week, and it deserves the second look.
The mechanism is a new LanguageModel protocol. Nearly any model can now back a LanguageModelSession behind one Swift API: Apple's own 20-billion-parameter on-device model, or a remote Claude, or Gemini. Anthropic and Google are both shipping Swift packages so developers call their models through the exact same interface they already use for on-device inference. Apple also added image input, a Python SDK, custom skills, and server-side execution, and made its newest cloud models free on Private Cloud Compute for small developers under two million downloads.
Said plainly, Apple is turning Apple Intelligence from a closed feature into an open developer platform. Write your app against one Swift interface, then swap the model behind it: on-device when you want privacy and zero cost, Claude when the reasoning is hard, all the same code. That's the agent-OS race from the opposite direction. HarmonyOS and Microsoft's MXC are about the OS becoming the agent. This is about the OS becoming the neutral runtime any model plugs into.
Who won? Developers, who stop being locked to one model, and Claude and Gemini, who just got a Swift import into every iPhone app. Who's exposed? Apple's own models, now competing head to head inside their own framework. The bet underneath is the same one Apple made by renting Gemini for Siri: owning the interface and the device beats owning the smartest model. One caveat, the open-source part is a pledge for later this summer, not a release yet, so watch whether it actually ships. Coverage from Apple's WWDC 2026 framework sessions.
← Back to all articles
The mechanism is a new LanguageModel protocol. Nearly any model can now back a LanguageModelSession behind one Swift API: Apple's own 20-billion-parameter on-device model, or a remote Claude, or Gemini. Anthropic and Google are both shipping Swift packages so developers call their models through the exact same interface they already use for on-device inference. Apple also added image input, a Python SDK, custom skills, and server-side execution, and made its newest cloud models free on Private Cloud Compute for small developers under two million downloads.
Said plainly, Apple is turning Apple Intelligence from a closed feature into an open developer platform. Write your app against one Swift interface, then swap the model behind it: on-device when you want privacy and zero cost, Claude when the reasoning is hard, all the same code. That's the agent-OS race from the opposite direction. HarmonyOS and Microsoft's MXC are about the OS becoming the agent. This is about the OS becoming the neutral runtime any model plugs into.
Who won? Developers, who stop being locked to one model, and Claude and Gemini, who just got a Swift import into every iPhone app. Who's exposed? Apple's own models, now competing head to head inside their own framework. The bet underneath is the same one Apple made by renting Gemini for Siri: owning the interface and the device beats owning the smartest model. One caveat, the open-source part is a pledge for later this summer, not a release yet, so watch whether it actually ships. Coverage from Apple's WWDC 2026 framework sessions.
Comments